H20s to China + 15% with Chris Miller and Lennart Heim

The former President’s off-the-cuff description of Nvidia’s H20 AI chip as “obsolete” hides a far more complex and strategically significant reality, one where the chip’s exceptional memory bandwidth makes it a potent tool for deploying advanced AI models, not just training them. This deep dive into the ongoing saga of U.S. export controls on AI semiconductors traces the H20’s journey from a banned product to a contentious commodity sold with a 15% “export fee,” framing a critical debate about technology, national security, and economic statecraft. Experts dissect whether selling such chips to China merely feeds a competitor or strategically maintains American leverage and delays the development of a domestic Chinese chip ecosystem.

Central to the discussion is a technical breakdown of what makes the H20 significant. While its raw computational power (FLOPS) is significantly lower than Nvidia’s top-tier chips like the H100 or the forthcoming Blackwell, its high-bandwidth memory (HBM) specs are actually superior. This makes it exceptionally good at “inference”—the process of running already-trained AI models to generate useful outputs—which is increasingly vital as AI systems require more “thinking” time per query. The argument that Huawei has a “similar chip” is countered by analyses of both quality and, more importantly, quantity: Huawei’s production capacity is orders of magnitude smaller than Nvidia’s, and its chips lack comparable memory performance and software maturity.

The conversation weighs two opposing strategic philosophies. One argues that restricting AI compute is essential to hobble a geopolitical rival’s progress in a domain that could redefine economic and military power. The other contends that flooding the Chinese market with U.S. chips retards domestic innovation, keeps Chinese developers dependent on the American software stack (like CUDA), and provides ongoing leverage and revenue. Skepticism is voiced about the “silicon shield” theory—that chip dependence deters a Taiwan conflict—with commentators suggesting broader geopolitical factors are far more decisive to Beijing’s calculus than access to a specific generation of AI accelerators.

Looking forward, the analysis extends beyond the H20 to the fraught political economy shaping these decisions. The personal, deal-centric approach of the administration raises questions about whether national security tools like export controls should be wielded as bargaining chips in trade negotiations. Furthermore, the reaction within China—with state media suddenly amplifying fears of “kill switches” in U.S. chips—highlights a mirroring internal debate. Different Chinese ministries and companies have competing interests: securing advanced technology, promoting self-sufficiency, and managing costs. The ultimate impact of the H20 decision will hinge on how these U.S. and Chinese political and corporate dynamics interact in the coming months.

Surprising Insights

• Memory Over Muscle: The supposedly “obsolete” H20 chip actually boasts better high-bandwidth memory performance (4 terabytes/second) than Nvidia’s flagship H100, making it particularly potent for the critical and growing task of AI inference and complex model “thinking.”
• The Quantity Chasm: Even optimistic estimates put Huawei’s 2025 production of its best AI chip (the Ascend 910C) at around 200,000 units. In contrast, Nvidia could sell over a million H20s into China, highlighting a production gap that market share percentages alone obscure.
• Security for Sale: The decision pathway—using a national security tool (export controls) to create a revenue stream (a 15% fee) without clear, public national security concessions from China—represents a novel and controversial blending of security and trade policy.
• Corporate Silence as Strategy: The lack of vocal opposition from most major U.S. AI companies (aside from those like Anthropic with alternative chip suppliers) is interpreted as a reflection of their continued dependence on Nvidia’s hardware, creating a muted political economy around the issue.

Practical Takeaways

• For policymakers, the debate suggests a potential middle path: prioritize selling controlled cloud access to AI capabilities over shipping physical chips, maintaining more direct oversight and immediate leverage while still capturing economic value.
• If the goal is to meaningfully slow a competitor’s advance, focus on restricting the enabling tools and materials (like High-Bandwidth Memory and semiconductor manufacturing equipment) rather than just the end-product chips, as this targets the root of their innovation capacity.
• When analyzing the AI chip landscape, look beyond headline FLOPS and deeply examine memory bandwidth, software ecosystem lock-in, and production scale, as these factors often determine real-world utility and strategic dependency.
• Track the licensing conditions and volume caps attached to any sales approvals, as a high volume of even moderately inferior chips can have a greater aggregate impact than a few units of a superior, banned chip.